[AMBARI-22644] Node Managers fail to start after Spark2 is patched due to CNF YarnShuffleService

• Type: Bug
• Status: Resolved
• Priority: Critical
• Resolution: Fixed
• Affects Version/s: 2.6.1
• Fix Version/s: 2.6.2
• Assignee: Jonathan Hurley
• Reporter: Vivek Sharma
• Created: 13/Dec/17 17:05
• Updated: 18/Jan/18 18:26
• Resolved: 22/Dec/17 14:52

Description

STR

1. Deploy HDP-2.6.4.0 cluster with Ambari-2.6.1.0-114
2. Apply HBase patch Upgrade on the cluster (this step is optional)
3. Then apply Spark2 patch Upgrade on the cluster
4. Restart Node Managers

Result
NM restart fails with below error:

2017-12-10 07:17:02,559 INFO  impl.MetricsSystemImpl (MetricsSystemImpl.java:shutdown(606)) - NodeManager metrics system shutdown complete.
2017-12-10 07:17:02,559 FATAL nodemanager.NodeManager (NodeManager.java:initAndStartNodeManager(549)) - Error starting NodeManager
org.apache.hadoop.service.ServiceStateException: java.lang.ClassNotFoundException: org.apache.spark.network.yarn.YarnShuffleService
        at org.apache.hadoop.service.ServiceStateException.convert(ServiceStateException.java:59)
        at org.apache.hadoop.service.AbstractService.init(AbstractService.java:172)
        at org.apache.hadoop.service.CompositeService.serviceInit(CompositeService.java:107)
        at org.apache.hadoop.yarn.server.nodemanager.containermanager.ContainerManagerImpl.serviceInit(ContainerManagerImpl.java:245)
        at org.apache.hadoop.service.AbstractService.init(AbstractService.java:163)
        at org.apache.hadoop.service.CompositeService.serviceInit(CompositeService.java:107)
        at org.apache.hadoop.yarn.server.nodemanager.NodeManager.serviceInit(NodeManager.java:291)
        at org.apache.hadoop.service.AbstractService.init(AbstractService.java:163)
        at org.apache.hadoop.yarn.server.nodemanager.NodeManager.initAndStartNodeManager(NodeManager.java:546)
        at org.apache.hadoop.yarn.server.nodemanager.NodeManager.main(NodeManager.java:594)
Caused by: java.lang.ClassNotFoundException: org.apache.spark.network.yarn.YarnShuffleService
        at java.net.URLClassLoader.findClass(URLClassLoader.java:381)
        at java.lang.ClassLoader.loadClass(ClassLoader.java:424)
        at sun.misc.Launcher$AppClassLoader.loadClass(Launcher.java:335)
        at java.lang.ClassLoader.loadClass(ClassLoader.java:357)
        at org.apache.hadoop.util.ApplicationClassLoader.loadClass(ApplicationClassLoader.java:197)
        at org.apache.hadoop.util.ApplicationClassLoader.loadClass(ApplicationClassLoader.java:165)
        at java.lang.Class.forName0(Native Method)
        at java.lang.Class.forName(Class.java:348)
        at org.apache.hadoop.yarn.server.nodemanager.containermanager.AuxiliaryServiceWithCustomClassLoader.getInstance(AuxiliaryServiceWithCustomClassLoader.java:169)
        at org.apache.hadoop.yarn.server.nodemanager.containermanager.AuxServices.serviceInit(AuxServices.java:131)
        at org.apache.hadoop.service.AbstractService.init(AbstractService.java:163)
        ... 8 more
2017-12-10 07:17:02,562 INFO  nodemanager.NodeManager (LogAdapter.java:info(45)) - SHUTDOWN_MSG:

The spark properties are correctly being written out as per AMBARI-22525.

Initially, we had defined Spark properties for ATS like this:

    <name>yarn.nodemanager.aux-services.spark_shuffle.classpath</name>
    <value>{{stack_root}}/${hdp.version}/spark/aux/*</value>

When YARN upgrades without Spark, we run into AMBARI-22525. Seems like the shuffle classes are installed as part of RPM dependencies, but not the SparkATSPlugin.

So:

• If we use YARN's version for the Spark classes, then ATS can't find SparkATSPlugin since that is not part of YARN.
• If we use Spark's version for the classes, then Spark can never upgrade without YARN since NodeManager can't find the new Spark classes.

However, it seems like shuffle and ATS use different properties. We changed all 3 properties in AMBARI-22525:

yarn.nodemanager.aux-services.spark2_shuffle.classpath
yarn.nodemanager.aux-services.spark_shuffle.classpath
yarn.timeline-service.entity-group-fs-store.group-id-plugin-classpath

It seems like what need to do is change the spark shuffle stuff back to hdp.version, but leave ATS using the new version since we're guaranteed to have Spark installed on the ATS machine.

Issue Links

• links to: ReviewBoard